Error audit

Skill christophecapel/claude-mechanisms-tools/skills/error-audit

Audit errors across Claude Code session transcripts. Scans ~/.claude/projects/*.jsonl for 7 error classes (tool_error, validation_error, permission_denial, hook_block, bash_fail, retry_storm, read_before_edit), clusters by root-cause signature, and surfaces top N with suggested remediation tiers. Use when the user says "error audit", "errors across sessions", "system health errors", or "/error-audit".From its SKILL.md

Install

npx -y skills add christophecapel/claude-mechanisms-tools --skill error-audit

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

1 stars1 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
runs commandsInstructs the agent to run 8 commands, including `python3 <skill-path>/error-audit.py` and 7 more.

SKILL.md

3.3 KB, 713 tokens by cl100k_base, as published. Nobody here has run it

/error-audit — Cross-session error audit

Scan every Claude Code session transcript for errors, cluster by root cause, surface the top offenders with suggested remediations.

Usage

# Default: all sessions, top 20 clusters, human output
python3 error-audit.py

# Last 30 days only
python3 error-audit.py --since 30

# Show fewer clusters
python3 error-audit.py --top 10

# Machine-readable (for piping into other tools)
python3 error-audit.py --json

# Override the projects dir (useful for testing)
python3 error-audit.py --projects-dir /path/to/projects

# Override or disable suppressions
python3 error-audit.py --suppressions-path /path/to/suppressions.md
python3 error-audit.py --no-suppressions
python3 error-audit.py --show-suppressed

Steps (when invoking via `/error-audit`)

Run python3 <skill-path>/error-audit.py with any arguments the user provided.
Display the output to the user exactly as printed (includes colour-coded counts and suggested remediation tiers).
For the top 3 clusters, propose a concrete action:
- Tier 1 (settings allowlist): show the exact ~/.claude/settings.json entry to add, but do NOT auto-apply — user must review Bash allowlist changes.
- Tier 2 (hook or script fix): locate the hook/script and propose the edit.
- Tier 3 (instruction/memory): draft the feedback or mechanism memory entry.
Do NOT auto-apply any remediation. This skill surfaces and proposes; the user decides.

Configuration

Env var	Purpose	Default
`CLAUDE_ERROR_AUDIT_SUPPRESSIONS`	Suppressions file path (cluster_keys to hide as working-as-designed)	`<skill-dir>/suppressions.md` (ships with toolkit)

The suppressions file ships with one entry by default: the plan-review-gate's intentional permission_denial:ExitPlanMode blocks. Add your own entries to silence known-good clusters.

What counts as "actionable"

A cluster is actionable if the signature is concrete enough to map to a single fix:

permission_denial:Bash:git → allowlist-adjacent, concrete
tool_error:Read:File does not exist: <X>/... → behavioural, concrete ("Glob before Read")
hook_block:Stop:response-linter:Want me to → specific linter rule, concrete

Clusters with vague signatures (single-count generic errors) are monitor-only until a pattern forms.

Implements

Mechanism #19 — Detection rules: more specific patterns, never broader allowlists
Mechanism #21 — Structural intervention beats pattern N+1

See https://github.com/christophecapel/claude-mechanisms

Out of scope (this release)

/error-audit-triage (interactive remediation flow with Linear ticket creation) — myOS-coupled today, deferred to v0.3.1+ once the agent path + Linear flow are decoupled.
error-audit-post.py (post findings to a GitHub health-check issue) — myOS-specific health-check format, stays myOS-only.

What ships with it: 2 files

22.8 KB alongside SKILL.md, 1 of them executable

error-audit.pyruns20.5 KB
suppressions.md2.3 KB